4 * Copyright (c) 2006 The DragonFly Project. All rights reserved.
5 * Copyright (c) 1991 Regents of the University of California.
7 * Copyright (c) 1994 John S. Dyson
9 * Copyright (c) 1994 David Greenman
10 * All rights reserved.
11 * Copyright (c) 2004-2006 Matthew Dillon
12 * All rights reserved.
14 * Redistribution and use in source and binary forms, with or without
15 * modification, are permitted provided that the following conditions
18 * 1. Redistributions of source code must retain the above copyright
19 * notice, this list of conditions and the following disclaimer.
20 * 2. Redistributions in binary form must reproduce the above copyright
21 * notice, this list of conditions and the following disclaimer in
22 * the documentation and/or other materials provided with the
24 * 3. Neither the name of The DragonFly Project nor the names of its
25 * contributors may be used to endorse or promote products derived
26 * from this software without specific, prior written permission.
28 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
29 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
30 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
31 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
32 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
33 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
34 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
35 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
36 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
37 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
38 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
41 * from: @(#)pmap.c 7.7 (Berkeley) 5/12/91
42 * $FreeBSD: src/sys/i386/i386/pmap.c,v 1.250.2.18 2002/03/06 22:48:53 silby Exp $
45 * NOTE: PMAP_INVAL_ADD: In pc32 this function is called prior to adjusting
46 * the PTE in the page table, because a cpu synchronization might be required.
47 * The actual invalidation is delayed until the following call or flush. In
48 * the VKERNEL build this function is called prior to adjusting the PTE and
49 * invalidates the table synchronously (not delayed), and is not SMP safe
53 #include <sys/types.h>
54 #include <sys/systm.h>
55 #include <sys/kernel.h>
58 #include <sys/vkernel.h>
60 #include <sys/thread.h>
62 #include <sys/vmspace.h>
65 #include <vm/vm_page.h>
66 #include <vm/vm_extern.h>
67 #include <vm/vm_kern.h>
68 #include <vm/vm_object.h>
69 #include <vm/vm_zone.h>
70 #include <vm/vm_pageout.h>
72 #include <machine/md_var.h>
73 #include <machine/pcb.h>
74 #include <machine/pmap_inval.h>
75 #include <machine/globaldata.h>
77 #include <sys/sysref2.h>
78 #include <sys/spinlock2.h>
82 struct pmap kernel_pmap;
84 static struct vm_zone pvzone;
85 static struct vm_object pvzone_obj;
86 static TAILQ_HEAD(,pmap) pmap_list = TAILQ_HEAD_INITIALIZER(pmap_list);
87 static int pv_entry_count;
88 static int pv_entry_max;
89 static int pv_entry_high_water;
90 static int pmap_pagedaemon_waken;
91 static boolean_t pmap_initialized = FALSE;
92 static int protection_codes[8];
94 static void i386_protection_init(void);
95 static void pmap_remove_all(vm_page_t m);
96 static int pmap_release_free_page(struct pmap *pmap, vm_page_t p);
99 #ifndef PMAP_SHPGPERPROC
100 #define PMAP_SHPGPERPROC 200
103 #define pmap_pde(m, v) (&((m)->pm_pdir[(vm_offset_t)(v) >> PDRSHIFT]))
105 #define pte_prot(m, p) \
106 (protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
112 struct pv_entry *pvinit;
114 for (i = 0; i < vm_page_array_size; i++) {
117 m = &vm_page_array[i];
118 TAILQ_INIT(&m->md.pv_list);
119 m->md.pv_list_count = 0;
122 i = vm_page_array_size;
125 pvinit = (struct pv_entry *)kmem_alloc(&kernel_map, i*sizeof(*pvinit));
126 zbootinit(&pvzone, "PV ENTRY", sizeof(*pvinit), pvinit, i);
127 pmap_initialized = TRUE;
133 int shpgperproc = PMAP_SHPGPERPROC;
135 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
136 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
137 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
138 pv_entry_high_water = 9 * (pv_entry_max / 10);
139 zinitna(&pvzone, &pvzone_obj, NULL, 0, pv_entry_max, ZONE_INTERRUPT, 1);
143 * Bootstrap the kernel_pmap so it can be used with pmap_enter().
145 * NOTE! pm_pdir for the kernel pmap is offset so VA's translate
146 * directly into PTD indexes (PTA is also offset for the same reason).
147 * This is necessary because, for now, KVA is not mapped at address 0.
149 * Page table pages are not managed like they are in normal pmaps, so
150 * no pteobj is needed.
155 vm_pindex_t i = (vm_offset_t)KernelPTD >> PAGE_SHIFT;
158 * The kernel_pmap's pm_pteobj is used only for locking and not
161 kernel_pmap.pm_pdir = KernelPTD - (KvaStart >> SEG_SHIFT);
162 kernel_pmap.pm_pdirpte = KernelPTA[i];
163 kernel_pmap.pm_count = 1;
164 kernel_pmap.pm_active = (cpumask_t)-1 & ~CPUMASK_LOCK;
165 kernel_pmap.pm_pteobj = &kernel_object;
166 TAILQ_INIT(&kernel_pmap.pm_pvlist);
167 TAILQ_INIT(&kernel_pmap.pm_pvlist_free);
168 spin_init(&kernel_pmap.pm_spin);
169 lwkt_token_init(&kernel_pmap.pm_token, "kpmap_tok");
170 i386_protection_init();
174 * Initialize pmap0/vmspace0 . Since process 0 never enters user mode we
175 * just dummy it up so it works well enough for fork().
177 * In DragonFly, process pmaps may only be used to manipulate user address
178 * space, never kernel address space.
181 pmap_pinit0(struct pmap *pmap)
186 /************************************************************************
187 * Procedures to manage whole physical maps *
188 ************************************************************************
190 * Initialize a preallocated and zeroed pmap structure,
191 * such as one in a vmspace structure.
194 pmap_pinit(struct pmap *pmap)
200 * No need to allocate page table space yet but we do need a valid
201 * page directory table.
203 if (pmap->pm_pdir == NULL) {
205 (vpte_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
209 * allocate object for the pte array and page directory
211 npages = VPTE_PAGETABLE_SIZE +
212 (VM_MAX_USER_ADDRESS / PAGE_SIZE) * sizeof(vpte_t);
213 npages = (npages + PAGE_MASK) / PAGE_SIZE;
215 if (pmap->pm_pteobj == NULL)
216 pmap->pm_pteobj = vm_object_allocate(OBJT_DEFAULT, npages);
217 pmap->pm_pdindex = npages - 1;
220 * allocate the page directory page
222 ptdpg = vm_page_grab(pmap->pm_pteobj, pmap->pm_pdindex,
223 VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
225 ptdpg->wire_count = 1;
226 atomic_add_int(&vmstats.v_wire_count, 1);
228 /* not usually mapped */
229 ptdpg->valid = VM_PAGE_BITS_ALL;
230 vm_page_flag_clear(ptdpg, PG_MAPPED);
231 vm_page_wakeup(ptdpg);
233 pmap_kenter((vm_offset_t)pmap->pm_pdir, VM_PAGE_TO_PHYS(ptdpg));
234 pmap->pm_pdirpte = KernelPTA[(vm_offset_t)pmap->pm_pdir >> PAGE_SHIFT];
235 if ((ptdpg->flags & PG_ZERO) == 0)
236 bzero(pmap->pm_pdir, PAGE_SIZE);
237 vm_page_flag_clear(ptdpg, PG_ZERO);
241 pmap->pm_ptphint = NULL;
242 pmap->pm_cpucachemask = 0;
243 TAILQ_INIT(&pmap->pm_pvlist);
244 TAILQ_INIT(&pmap->pm_pvlist_free);
245 spin_init(&pmap->pm_spin);
246 lwkt_token_init(&pmap->pm_token, "pmap_tok");
247 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
248 pmap->pm_stats.resident_count = 1;
252 * Clean up a pmap structure so it can be physically freed
257 pmap_puninit(pmap_t pmap)
260 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pdir, PAGE_SIZE);
261 pmap->pm_pdir = NULL;
263 if (pmap->pm_pteobj) {
264 vm_object_deallocate(pmap->pm_pteobj);
265 pmap->pm_pteobj = NULL;
271 * Wire in kernel global address entries. To avoid a race condition
272 * between pmap initialization and pmap_growkernel, this procedure
273 * adds the pmap to the master list (which growkernel scans to update),
274 * then copies the template.
276 * In a virtual kernel there are no kernel global address entries.
281 pmap_pinit2(struct pmap *pmap)
283 spin_lock(&pmap_spin);
284 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
285 spin_unlock(&pmap_spin);
289 * Release all resources held by the given physical map.
291 * Should only be called if the map contains no valid mappings.
293 * Caller must hold pmap->pm_token
295 static int pmap_release_callback(struct vm_page *p, void *data);
298 pmap_release(struct pmap *pmap)
300 struct mdglobaldata *gd = mdcpu;
301 vm_object_t object = pmap->pm_pteobj;
302 struct rb_vm_page_scan_info info;
304 KKASSERT(pmap != &kernel_pmap);
306 #if defined(DIAGNOSTIC)
307 if (object->ref_count != 1)
308 panic("pmap_release: pteobj reference count != 1");
311 * Once we destroy the page table, the mapping becomes invalid.
312 * Don't waste time doing a madvise to invalidate the mapping, just
313 * set cpucachemask to 0.
315 if (pmap->pm_pdir == gd->gd_PT1pdir) {
316 gd->gd_PT1pdir = NULL;
318 /* madvise(gd->gd_PT1map, SEG_SIZE, MADV_INVAL); */
320 if (pmap->pm_pdir == gd->gd_PT2pdir) {
321 gd->gd_PT2pdir = NULL;
323 /* madvise(gd->gd_PT2map, SEG_SIZE, MADV_INVAL); */
325 if (pmap->pm_pdir == gd->gd_PT3pdir) {
326 gd->gd_PT3pdir = NULL;
328 /* madvise(gd->gd_PT3map, SEG_SIZE, MADV_INVAL); */
332 info.object = object;
334 spin_lock(&pmap_spin);
335 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
336 spin_unlock(&pmap_spin);
338 vm_object_hold(object);
342 info.limit = object->generation;
344 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
345 pmap_release_callback, &info);
346 if (info.error == 0 && info.mpte) {
347 if (!pmap_release_free_page(pmap, info.mpte))
350 } while (info.error);
351 vm_object_drop(object);
354 * Leave the KVA reservation for pm_pdir cached for later reuse.
356 pmap->pm_pdirpte = 0;
357 pmap->pm_cpucachemask = 0;
361 * Callback to release a page table page backing a directory
365 pmap_release_callback(struct vm_page *p, void *data)
367 struct rb_vm_page_scan_info *info = data;
369 if (p->pindex == info->pmap->pm_pdindex) {
373 if (!pmap_release_free_page(info->pmap, p)) {
377 if (info->object->generation != info->limit) {
385 * Retire the given physical map from service. Should only be called if
386 * the map contains no valid mappings.
391 pmap_destroy(pmap_t pmap)
396 lwkt_gettoken(&vm_token);
397 if (--pmap->pm_count == 0) {
399 panic("destroying a pmap is not yet implemented");
401 lwkt_reltoken(&vm_token);
405 * Add a reference to the specified pmap.
410 pmap_reference(pmap_t pmap)
413 lwkt_gettoken(&vm_token);
415 lwkt_reltoken(&vm_token);
419 /************************************************************************
420 * VMSPACE MANAGEMENT *
421 ************************************************************************
423 * The VMSPACE management we do in our virtual kernel must be reflected
424 * in the real kernel. This is accomplished by making vmspace system
425 * calls to the real kernel.
428 cpu_vmspace_alloc(struct vmspace *vm)
433 #define LAST_EXTENT (VM_MAX_USER_ADDRESS - 0x80000000)
435 if (vmspace_create(&vm->vm_pmap, 0, NULL) < 0)
436 panic("vmspace_create() failed");
438 rp = vmspace_mmap(&vm->vm_pmap, (void *)0x00000000, 0x40000000,
439 PROT_READ|PROT_WRITE,
440 MAP_FILE|MAP_SHARED|MAP_VPAGETABLE|MAP_FIXED,
442 if (rp == MAP_FAILED)
443 panic("vmspace_mmap: failed1");
444 vmspace_mcontrol(&vm->vm_pmap, (void *)0x00000000, 0x40000000,
446 rp = vmspace_mmap(&vm->vm_pmap, (void *)0x40000000, 0x40000000,
447 PROT_READ|PROT_WRITE,
448 MAP_FILE|MAP_SHARED|MAP_VPAGETABLE|MAP_FIXED,
449 MemImageFd, 0x40000000);
450 if (rp == MAP_FAILED)
451 panic("vmspace_mmap: failed2");
452 vmspace_mcontrol(&vm->vm_pmap, (void *)0x40000000, 0x40000000,
454 rp = vmspace_mmap(&vm->vm_pmap, (void *)0x80000000, LAST_EXTENT,
455 PROT_READ|PROT_WRITE,
456 MAP_FILE|MAP_SHARED|MAP_VPAGETABLE|MAP_FIXED,
457 MemImageFd, 0x80000000);
458 vmspace_mcontrol(&vm->vm_pmap, (void *)0x80000000, LAST_EXTENT,
460 if (rp == MAP_FAILED)
461 panic("vmspace_mmap: failed3");
463 r = vmspace_mcontrol(&vm->vm_pmap, (void *)0x00000000, 0x40000000,
464 MADV_SETMAP, vmspace_pmap(vm)->pm_pdirpte);
466 panic("vmspace_mcontrol: failed1");
467 r = vmspace_mcontrol(&vm->vm_pmap, (void *)0x40000000, 0x40000000,
468 MADV_SETMAP, vmspace_pmap(vm)->pm_pdirpte);
470 panic("vmspace_mcontrol: failed2");
471 r = vmspace_mcontrol(&vm->vm_pmap, (void *)0x80000000, LAST_EXTENT,
472 MADV_SETMAP, vmspace_pmap(vm)->pm_pdirpte);
474 panic("vmspace_mcontrol: failed3");
478 cpu_vmspace_free(struct vmspace *vm)
480 if (vmspace_destroy(&vm->vm_pmap) < 0)
481 panic("vmspace_destroy() failed");
484 /************************************************************************
485 * Procedures which operate directly on the kernel PMAP *
486 ************************************************************************/
489 * This maps the requested page table and gives us access to it.
491 * This routine can be called from a potentially preempting interrupt
492 * thread or from a normal thread.
495 get_ptbase(struct pmap *pmap, vm_offset_t va)
497 struct mdglobaldata *gd = mdcpu;
499 if (pmap == &kernel_pmap) {
500 KKASSERT(va >= KvaStart && va < KvaEnd);
501 return(KernelPTA + (va >> PAGE_SHIFT));
502 } else if (pmap->pm_pdir == gd->gd_PT1pdir) {
503 if ((pmap->pm_cpucachemask & gd->mi.gd_cpumask) == 0) {
504 *gd->gd_PT1pde = pmap->pm_pdirpte;
505 madvise(gd->gd_PT1map, SEG_SIZE, MADV_INVAL);
506 atomic_set_cpumask(&pmap->pm_cpucachemask,
509 return(gd->gd_PT1map + (va >> PAGE_SHIFT));
510 } else if (pmap->pm_pdir == gd->gd_PT2pdir) {
511 if ((pmap->pm_cpucachemask & gd->mi.gd_cpumask) == 0) {
512 *gd->gd_PT2pde = pmap->pm_pdirpte;
513 madvise(gd->gd_PT2map, SEG_SIZE, MADV_INVAL);
514 atomic_set_cpumask(&pmap->pm_cpucachemask,
517 return(gd->gd_PT2map + (va >> PAGE_SHIFT));
521 * If we aren't running from a potentially preempting interrupt,
522 * load a new page table directory into the page table cache
524 if (gd->mi.gd_intr_nesting_level == 0 &&
525 (gd->mi.gd_curthread->td_flags & TDF_INTTHREAD) == 0) {
527 * Choose one or the other and map the page table
528 * in the KVA space reserved for it.
530 if ((gd->gd_PTflip = 1 - gd->gd_PTflip) == 0) {
531 gd->gd_PT1pdir = pmap->pm_pdir;
532 *gd->gd_PT1pde = pmap->pm_pdirpte;
533 madvise(gd->gd_PT1map, SEG_SIZE, MADV_INVAL);
534 atomic_set_cpumask(&pmap->pm_cpucachemask,
536 return(gd->gd_PT1map + (va >> PAGE_SHIFT));
538 gd->gd_PT2pdir = pmap->pm_pdir;
539 *gd->gd_PT2pde = pmap->pm_pdirpte;
540 madvise(gd->gd_PT2map, SEG_SIZE, MADV_INVAL);
541 atomic_set_cpumask(&pmap->pm_cpucachemask,
543 return(gd->gd_PT2map + (va >> PAGE_SHIFT));
548 * If we are running from a preempting interrupt use a private
549 * map. The caller must be in a critical section.
551 KKASSERT(IN_CRITICAL_SECT(curthread));
552 if (pmap->pm_pdir == gd->gd_PT3pdir) {
553 if ((pmap->pm_cpucachemask & gd->mi.gd_cpumask) == 0) {
554 *gd->gd_PT3pde = pmap->pm_pdirpte;
555 madvise(gd->gd_PT3map, SEG_SIZE, MADV_INVAL);
556 atomic_set_cpumask(&pmap->pm_cpucachemask,
560 gd->gd_PT3pdir = pmap->pm_pdir;
561 *gd->gd_PT3pde = pmap->pm_pdirpte;
562 madvise(gd->gd_PT3map, SEG_SIZE, MADV_INVAL);
563 atomic_set_cpumask(&pmap->pm_cpucachemask,
566 return(gd->gd_PT3map + (va >> PAGE_SHIFT));
570 get_ptbase1(struct pmap *pmap, vm_offset_t va)
572 struct mdglobaldata *gd = mdcpu;
574 if (pmap == &kernel_pmap) {
575 KKASSERT(va >= KvaStart && va < KvaEnd);
576 return(KernelPTA + (va >> PAGE_SHIFT));
577 } else if (pmap->pm_pdir == gd->gd_PT1pdir) {
578 if ((pmap->pm_cpucachemask & gd->mi.gd_cpumask) == 0) {
579 *gd->gd_PT1pde = pmap->pm_pdirpte;
580 madvise(gd->gd_PT1map, SEG_SIZE, MADV_INVAL);
581 atomic_set_cpumask(&pmap->pm_cpucachemask,
584 return(gd->gd_PT1map + (va >> PAGE_SHIFT));
586 KKASSERT(gd->mi.gd_intr_nesting_level == 0 &&
587 (gd->mi.gd_curthread->td_flags & TDF_INTTHREAD) == 0);
588 gd->gd_PT1pdir = pmap->pm_pdir;
589 *gd->gd_PT1pde = pmap->pm_pdirpte;
590 madvise(gd->gd_PT1map, SEG_SIZE, MADV_INVAL);
591 return(gd->gd_PT1map + (va >> PAGE_SHIFT));
595 get_ptbase2(struct pmap *pmap, vm_offset_t va)
597 struct mdglobaldata *gd = mdcpu;
599 if (pmap == &kernel_pmap) {
600 KKASSERT(va >= KvaStart && va < KvaEnd);
601 return(KernelPTA + (va >> PAGE_SHIFT));
602 } else if (pmap->pm_pdir == gd->gd_PT2pdir) {
603 if ((pmap->pm_cpucachemask & gd->mi.gd_cpumask) == 0) {
604 *gd->gd_PT2pde = pmap->pm_pdirpte;
605 madvise(gd->gd_PT2map, SEG_SIZE, MADV_INVAL);
606 atomic_set_cpumask(&pmap->pm_cpucachemask,
609 return(gd->gd_PT2map + (va >> PAGE_SHIFT));
611 KKASSERT(gd->mi.gd_intr_nesting_level == 0 &&
612 (gd->mi.gd_curthread->td_flags & TDF_INTTHREAD) == 0);
613 gd->gd_PT2pdir = pmap->pm_pdir;
614 *gd->gd_PT2pde = pmap->pm_pdirpte;
615 madvise(gd->gd_PT2map, SEG_SIZE, MADV_INVAL);
616 return(gd->gd_PT2map + (va >> PAGE_SHIFT));
620 * Return a pointer to the page table entry for the specified va in the
621 * specified pmap. NULL is returned if there is no valid page table page
624 static __inline vpte_t *
625 pmap_pte(struct pmap *pmap, vm_offset_t va)
629 ptep = &pmap->pm_pdir[va >> SEG_SHIFT];
633 return (get_ptbase(pmap, va));
639 * Enter a mapping into kernel_pmap. Mappings created in this fashion
640 * are not managed. Mappings must be immediately accessible on all cpus.
642 * Call pmap_inval_pte() to invalidate the virtual pte and clean out the
643 * real pmap and handle related races before storing the new vpte.
646 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
651 KKASSERT(va >= KvaStart && va < KvaEnd);
652 npte = (vpte_t)pa | VPTE_R | VPTE_W | VPTE_V;
653 ptep = KernelPTA + (va >> PAGE_SHIFT);
655 pmap_inval_pte(ptep, &kernel_pmap, va);
660 * Synchronize a kvm mapping originally made for the private use on
661 * some other cpu so it can be used on all cpus.
663 * XXX add MADV_RESYNC to improve performance.
666 pmap_kenter_sync(vm_offset_t va)
668 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
672 * Synchronize a kvm mapping originally made for the private use on
673 * some other cpu so it can be used on our cpu. Turns out to be the
674 * same madvise() call, because we have to sync the real pmaps anyway.
676 * XXX add MADV_RESYNC to improve performance.
679 pmap_kenter_sync_quick(vm_offset_t va)
681 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
686 * Make a previously read-only kernel mapping R+W (not implemented by
690 pmap_kmodify_rw(vm_offset_t va)
692 *pmap_kpte(va) |= VPTE_R | VPTE_W;
693 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
697 * Make a kernel mapping non-cacheable (not applicable to virtual kernels)
700 pmap_kmodify_nc(vm_offset_t va)
702 *pmap_kpte(va) |= VPTE_N;
703 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
709 * Map a contiguous range of physical memory to a KVM
712 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
714 vm_offset_t sva, virt;
717 while (start < end) {
718 pmap_kenter(virt, start);
727 pmap_kpte(vm_offset_t va)
731 KKASSERT(va >= KvaStart && va < KvaEnd);
732 ptep = KernelPTA + (va >> PAGE_SHIFT);
737 * Enter an unmanaged KVA mapping for the private use of the current
738 * cpu only. pmap_kenter_sync() may be called to make the mapping usable
741 * It is illegal for the mapping to be accessed by other cpus unleess
742 * pmap_kenter_sync*() is called.
745 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
750 KKASSERT(va >= KvaStart && va < KvaEnd);
752 npte = (vpte_t)pa | VPTE_R | VPTE_W | VPTE_V;
753 ptep = KernelPTA + (va >> PAGE_SHIFT);
755 pmap_inval_pte_quick(ptep, &kernel_pmap, va);
760 * Make a temporary mapping for a physical address. This is only intended
761 * to be used for panic dumps.
763 * The caller is responsible for calling smp_invltlb().
766 pmap_kenter_temporary(vm_paddr_t pa, long i)
768 pmap_kenter_quick(crashdumpmap + (i * PAGE_SIZE), pa);
769 return ((void *)crashdumpmap);
773 * Remove an unmanaged mapping created with pmap_kenter*().
776 pmap_kremove(vm_offset_t va)
780 KKASSERT(va >= KvaStart && va < KvaEnd);
782 ptep = KernelPTA + (va >> PAGE_SHIFT);
784 pmap_inval_pte(ptep, &kernel_pmap, va);
789 * Remove an unmanaged mapping created with pmap_kenter*() but synchronize
790 * only with this cpu.
792 * Unfortunately because we optimize new entries by testing VPTE_V later
793 * on, we actually still have to synchronize with all the cpus. XXX maybe
794 * store a junk value and test against 0 in the other places instead?
797 pmap_kremove_quick(vm_offset_t va)
801 KKASSERT(va >= KvaStart && va < KvaEnd);
803 ptep = KernelPTA + (va >> PAGE_SHIFT);
805 pmap_inval_pte(ptep, &kernel_pmap, va); /* NOT _quick */
810 * Extract the physical address from the kernel_pmap that is associated
811 * with the specified virtual address.
814 pmap_kextract(vm_offset_t va)
819 KKASSERT(va >= KvaStart && va < KvaEnd);
821 ptep = KernelPTA + (va >> PAGE_SHIFT);
822 pa = (vm_paddr_t)(*ptep & VPTE_FRAME) | (va & PAGE_MASK);
827 * Map a set of unmanaged VM pages into KVM.
830 pmap_qenter(vm_offset_t va, struct vm_page **m, int count)
832 KKASSERT(va >= KvaStart && va + count * PAGE_SIZE < KvaEnd);
836 ptep = KernelPTA + (va >> PAGE_SHIFT);
838 pmap_inval_pte(ptep, &kernel_pmap, va);
839 *ptep = (vpte_t)(*m)->phys_addr | VPTE_R | VPTE_W | VPTE_V;
847 * Undo the effects of pmap_qenter*().
850 pmap_qremove(vm_offset_t va, int count)
852 KKASSERT(va >= KvaStart && va + count * PAGE_SIZE < KvaEnd);
856 ptep = KernelPTA + (va >> PAGE_SHIFT);
858 pmap_inval_pte(ptep, &kernel_pmap, va);
865 /************************************************************************
866 * Misc support glue called by machine independant code *
867 ************************************************************************
869 * These routines are called by machine independant code to operate on
870 * certain machine-dependant aspects of processes, threads, and pmaps.
874 * Initialize MD portions of the thread structure.
877 pmap_init_thread(thread_t td)
879 /* enforce pcb placement */
880 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
881 td->td_savefpu = &td->td_pcb->pcb_save;
882 td->td_sp = (char *)td->td_pcb - 16;
886 * This routine directly affects the fork perf for a process.
889 pmap_init_proc(struct proc *p)
894 * Destroy the UPAGES for a process that has exited and disassociate
895 * the process from its thread.
898 pmap_dispose_proc(struct proc *p)
900 KASSERT(p->p_lock == 0, ("attempt to dispose referenced proc! %p", p));
904 * We pre-allocate all page table pages for kernel virtual memory so
905 * this routine will only be called if KVM has been exhausted.
910 pmap_growkernel(vm_offset_t kstart, vm_offset_t kend)
914 addr = (kend + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
916 lwkt_gettoken(&vm_token);
917 if (addr > virtual_end - SEG_SIZE)
918 panic("KVM exhausted");
919 kernel_vm_end = addr;
920 lwkt_reltoken(&vm_token);
924 * The modification bit is not tracked for any pages in this range. XXX
925 * such pages in this maps should always use pmap_k*() functions and not
928 * XXX User and kernel address spaces are independant for virtual kernels,
929 * this function only applies to the kernel pmap.
932 pmap_track_modified(pmap_t pmap, vm_offset_t va)
934 if (pmap != &kernel_pmap)
936 if ((va < clean_sva) || (va >= clean_eva))
942 /************************************************************************
943 * Procedures supporting managed page table pages *
944 ************************************************************************
946 * These procedures are used to track managed page table pages. These pages
947 * use the page table page's vm_page_t to track PTEs in the page. The
948 * page table pages themselves are arranged in a VM object, pmap->pm_pteobj.
950 * This allows the system to throw away page table pages for user processes
951 * at will and reinstantiate them on demand.
955 * This routine works like vm_page_lookup() but also blocks as long as the
956 * page is busy. This routine does not busy the page it returns.
958 * Unless the caller is managing objects whos pages are in a known state,
959 * the call should be made with a critical section held so the page's object
960 * association remains valid on return.
963 pmap_page_lookup(vm_object_t object, vm_pindex_t pindex)
967 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
968 m = vm_page_lookup_busy_wait(object, pindex, FALSE, "pplookp");
974 * This routine unholds page table pages, and if the hold count
975 * drops to zero, then it decrements the wire count.
977 * We must recheck that this is the last hold reference after busy-sleeping
981 _pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m)
983 vm_page_busy_wait(m, FALSE, "pmuwpt");
984 KASSERT(m->queue == PQ_NONE,
985 ("_pmap_unwire_pte_hold: %p->queue != PQ_NONE", m));
987 if (m->hold_count == 1) {
989 * Unmap the page table page.
991 KKASSERT(pmap->pm_pdir[m->pindex] != 0);
992 pmap_inval_pde(&pmap->pm_pdir[m->pindex], pmap,
993 (vm_offset_t)m->pindex << SEG_SHIFT);
994 KKASSERT(pmap->pm_stats.resident_count > 0);
995 --pmap->pm_stats.resident_count;
997 if (pmap->pm_ptphint == m)
998 pmap->pm_ptphint = NULL;
1001 * This was our last hold, the page had better be unwired
1002 * after we decrement wire_count.
1004 * FUTURE NOTE: shared page directory page could result in
1005 * multiple wire counts.
1009 KKASSERT(m->wire_count == 0);
1010 atomic_add_int(&vmstats.v_wire_count, -1);
1011 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1013 vm_page_free_zero(m);
1016 KKASSERT(m->hold_count > 1);
1024 pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m)
1026 KKASSERT(m->hold_count > 0);
1027 if (m->hold_count > 1) {
1031 return _pmap_unwire_pte_hold(pmap, m);
1036 * After removing a page table entry, this routine is used to
1037 * conditionally free the page, and manage the hold/wire counts.
1040 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte)
1044 ASSERT_LWKT_TOKEN_HELD(vm_object_token(pmap->pm_pteobj));
1048 * page table pages in the kernel_pmap are not managed.
1050 if (pmap == &kernel_pmap)
1052 ptepindex = (va >> PDRSHIFT);
1053 if (pmap->pm_ptphint &&
1054 (pmap->pm_ptphint->pindex == ptepindex)) {
1055 mpte = pmap->pm_ptphint;
1057 mpte = pmap_page_lookup(pmap->pm_pteobj, ptepindex);
1058 pmap->pm_ptphint = mpte;
1059 vm_page_wakeup(mpte);
1062 return pmap_unwire_pte_hold(pmap, mpte);
1066 * Attempt to release and free the vm_page backing a page directory page
1067 * in a pmap. Returns 1 on success, 0 on failure (if the procedure had
1071 pmap_release_free_page(struct pmap *pmap, vm_page_t p)
1073 vpte_t *pde = pmap->pm_pdir;
1076 * This code optimizes the case of freeing non-busy
1077 * page-table pages. Those pages are zero now, and
1078 * might as well be placed directly into the zero queue.
1080 if (vm_page_busy_try(p, FALSE)) {
1081 vm_page_sleep_busy(p, FALSE, "pmaprl");
1084 KKASSERT(pmap->pm_stats.resident_count > 0);
1085 --pmap->pm_stats.resident_count;
1087 if (p->hold_count) {
1088 panic("pmap_release: freeing held page table page");
1091 * Page directory pages need to have the kernel stuff cleared, so
1092 * they can go into the zero queue also.
1094 * In virtual kernels there is no 'kernel stuff'. For the moment
1095 * I just make sure the whole thing has been zero'd even though
1096 * it should already be completely zero'd.
1098 * pmaps for vkernels do not self-map because they do not share
1099 * their address space with the vkernel. Clearing of pde[] thus
1100 * only applies to page table pages and not to the page directory
1103 if (p->pindex == pmap->pm_pdindex) {
1104 bzero(pde, VPTE_PAGETABLE_SIZE);
1105 pmap_kremove((vm_offset_t)pmap->pm_pdir);
1107 KKASSERT(pde[p->pindex] != 0);
1108 pmap_inval_pde(&pde[p->pindex], pmap,
1109 (vm_offset_t)p->pindex << SEG_SHIFT);
1113 * Clear the matching hint
1115 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex))
1116 pmap->pm_ptphint = NULL;
1119 * And throw the page away. The page is completely zero'd out so
1120 * optimize the free call.
1123 atomic_add_int(&vmstats.v_wire_count, -1);
1124 vm_page_free_zero(p);
1129 * This routine is called if the page table page is not mapped in the page
1132 * The routine is broken up into two parts for readability.
1134 * It must return a held mpte and map the page directory page as required.
1135 * Because vm_page_grab() can block, we must re-check pm_pdir[ptepindex]
1138 _pmap_allocpte(pmap_t pmap, unsigned ptepindex)
1144 * Find or fabricate a new pagetable page. A busied page will be
1145 * returned. This call may block.
1147 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
1148 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
1150 if (m->valid == 0) {
1151 if ((m->flags & PG_ZERO) == 0)
1152 pmap_zero_page(VM_PAGE_TO_PHYS(m));
1153 m->valid = VM_PAGE_BITS_ALL;
1154 vm_page_flag_clear(m, PG_ZERO);
1156 KKASSERT((m->flags & PG_ZERO) == 0);
1158 vm_page_flag_set(m, PG_MAPPED);
1160 KASSERT(m->queue == PQ_NONE,
1161 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
1164 * Increment the hold count for the page we will be returning to
1170 * It is possible that someone else got in and mapped by the page
1171 * directory page while we were blocked, if so just unbusy and
1172 * return the held page.
1174 if ((ptepa = pmap->pm_pdir[ptepindex]) != 0) {
1175 KKASSERT((ptepa & VPTE_FRAME) == VM_PAGE_TO_PHYS(m));
1180 if (m->wire_count == 0)
1181 atomic_add_int(&vmstats.v_wire_count, 1);
1185 * Map the pagetable page into the process address space, if
1186 * it isn't already there.
1188 ++pmap->pm_stats.resident_count;
1190 ptepa = VM_PAGE_TO_PHYS(m);
1191 pmap->pm_pdir[ptepindex] = (vpte_t)ptepa | VPTE_R | VPTE_W | VPTE_V |
1195 * We are likely about to access this page table page, so set the
1196 * page table hint to reduce overhead.
1198 pmap->pm_ptphint = m;
1206 * Determine the page table page required to access the VA in the pmap
1207 * and allocate it if necessary. Return a held vm_page_t for the page.
1209 * Only used with user pmaps.
1212 pmap_allocpte(pmap_t pmap, vm_offset_t va)
1218 ASSERT_LWKT_TOKEN_HELD(vm_object_token(pmap->pm_pteobj));
1221 * Calculate pagetable page index
1223 ptepindex = va >> PDRSHIFT;
1226 * Get the page directory entry
1228 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
1231 * This supports switching from a 4MB page to a
1234 if (ptepa & VPTE_PS) {
1235 KKASSERT(pmap->pm_pdir[ptepindex] != 0);
1236 pmap_inval_pde(&pmap->pm_pdir[ptepindex], pmap,
1237 (vm_offset_t)ptepindex << SEG_SHIFT);
1242 * If the page table page is mapped, we just increment the
1243 * hold count, and activate it.
1247 * In order to get the page table page, try the
1250 if (pmap->pm_ptphint &&
1251 (pmap->pm_ptphint->pindex == ptepindex)) {
1252 m = pmap->pm_ptphint;
1254 m = pmap_page_lookup(pmap->pm_pteobj, ptepindex);
1255 pmap->pm_ptphint = m;
1262 * Here if the pte page isn't mapped, or if it has been deallocated.
1264 return _pmap_allocpte(pmap, ptepindex);
1267 /************************************************************************
1268 * Managed pages in pmaps *
1269 ************************************************************************
1271 * All pages entered into user pmaps and some pages entered into the kernel
1272 * pmap are managed, meaning that pmap_protect() and other related management
1273 * functions work on these pages.
1277 * free the pv_entry back to the free list. This function may be
1278 * called from an interrupt.
1280 static __inline void
1281 free_pv_entry(pv_entry_t pv)
1288 * get a new pv_entry, allocating a block from the system
1289 * when needed. This function may be called from an interrupt.
1295 if (pv_entry_high_water &&
1296 (pv_entry_count > pv_entry_high_water) &&
1297 (pmap_pagedaemon_waken == 0)) {
1298 pmap_pagedaemon_waken = 1;
1299 wakeup (&vm_pages_needed);
1301 return zalloc(&pvzone);
1305 * This routine is very drastic, but can save the system
1315 static int warningdone=0;
1317 if (pmap_pagedaemon_waken == 0)
1319 lwkt_gettoken(&vm_token);
1320 pmap_pagedaemon_waken = 0;
1322 if (warningdone < 5) {
1323 kprintf("pmap_collect: collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
1327 for (i = 0; i < vm_page_array_size; i++) {
1328 m = &vm_page_array[i];
1329 if (m->wire_count || m->hold_count)
1331 if (vm_page_busy_try(m, TRUE) == 0) {
1332 if (m->wire_count == 0 && m->hold_count == 0) {
1338 lwkt_reltoken(&vm_token);
1342 * If it is the first entry on the list, it is actually
1343 * in the header and we must copy the following entry up
1344 * to the header. Otherwise we must search the list for
1345 * the entry. In either case we free the now unused entry.
1347 * caller must hold vm_token
1350 pmap_remove_entry(struct pmap *pmap, vm_page_t m, vm_offset_t va)
1356 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
1357 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
1358 if (pmap == pv->pv_pmap && va == pv->pv_va)
1362 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
1363 if (va == pv->pv_va)
1369 * Note that pv_ptem is NULL if the page table page itself is not
1370 * managed, even if the page being removed IS managed.
1374 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1375 m->md.pv_list_count--;
1376 atomic_add_int(&m->object->agg_pv_list_count, -1);
1377 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
1378 if (TAILQ_EMPTY(&m->md.pv_list))
1379 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1380 ++pmap->pm_generation;
1381 vm_object_hold(pmap->pm_pteobj);
1382 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem);
1383 vm_object_drop(pmap->pm_pteobj);
1391 * Create a pv entry for page at pa for (pmap, va). If the page table page
1392 * holding the VA is managed, mpte will be non-NULL.
1395 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m)
1400 pv = get_pv_entry();
1405 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
1406 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1407 ++pmap->pm_generation;
1408 m->md.pv_list_count++;
1409 atomic_add_int(&m->object->agg_pv_list_count, 1);
1415 * pmap_remove_pte: do the things to unmap a page in a process
1418 pmap_remove_pte(struct pmap *pmap, vpte_t *ptq, vm_offset_t va)
1423 oldpte = pmap_inval_loadandclear(ptq, pmap, va);
1424 if (oldpte & VPTE_WIRED)
1425 --pmap->pm_stats.wired_count;
1426 KKASSERT(pmap->pm_stats.wired_count >= 0);
1430 * Machines that don't support invlpg, also don't support
1431 * VPTE_G. XXX VPTE_G is disabled for SMP so don't worry about
1434 if (oldpte & VPTE_G)
1435 madvise((void *)va, PAGE_SIZE, MADV_INVAL);
1437 KKASSERT(pmap->pm_stats.resident_count > 0);
1438 --pmap->pm_stats.resident_count;
1439 if (oldpte & VPTE_MANAGED) {
1440 m = PHYS_TO_VM_PAGE(oldpte);
1441 if (oldpte & VPTE_M) {
1442 #if defined(PMAP_DIAGNOSTIC)
1443 if (pmap_nw_modified((pt_entry_t) oldpte)) {
1445 "pmap_remove: modified page not writable: va: 0x%x, pte: 0x%x\n",
1449 if (pmap_track_modified(pmap, va))
1452 if (oldpte & VPTE_A)
1453 vm_page_flag_set(m, PG_REFERENCED);
1454 return pmap_remove_entry(pmap, m, va);
1456 return pmap_unuse_pt(pmap, va, NULL);
1465 * Remove a single page from a process address space.
1467 * This function may not be called from an interrupt if the pmap is
1471 pmap_remove_page(struct pmap *pmap, vm_offset_t va)
1476 * if there is no pte for this address, just skip it!!! Otherwise
1477 * get a local va for mappings for this pmap and remove the entry.
1479 if (*pmap_pde(pmap, va) != 0) {
1480 ptq = get_ptbase(pmap, va);
1482 pmap_remove_pte(pmap, ptq, va);
1488 * Remove the given range of addresses from the specified map.
1490 * It is assumed that the start and end are properly rounded to the
1493 * This function may not be called from an interrupt if the pmap is
1499 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
1503 vm_offset_t ptpaddr;
1504 vm_pindex_t sindex, eindex;
1509 vm_object_hold(pmap->pm_pteobj);
1510 lwkt_gettoken(&vm_token);
1511 KKASSERT(pmap->pm_stats.resident_count >= 0);
1512 if (pmap->pm_stats.resident_count == 0) {
1513 lwkt_reltoken(&vm_token);
1514 vm_object_drop(pmap->pm_pteobj);
1519 * special handling of removing one page. a very
1520 * common operation and easy to short circuit some
1523 if (((sva + PAGE_SIZE) == eva) &&
1524 ((pmap->pm_pdir[(sva >> PDRSHIFT)] & VPTE_PS) == 0)) {
1525 pmap_remove_page(pmap, sva);
1526 lwkt_reltoken(&vm_token);
1527 vm_object_drop(pmap->pm_pteobj);
1532 * Get a local virtual address for the mappings that are being
1535 * XXX this is really messy because the kernel pmap is not relative
1538 sindex = (sva >> PAGE_SHIFT);
1539 eindex = (eva >> PAGE_SHIFT);
1541 for (; sindex < eindex; sindex = pdnxt) {
1545 * Calculate index for next page table.
1547 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1548 if (pmap->pm_stats.resident_count == 0)
1551 pdirindex = sindex / NPDEPG;
1552 if (((ptpaddr = pmap->pm_pdir[pdirindex]) & VPTE_PS) != 0) {
1553 KKASSERT(pmap->pm_pdir[pdirindex] != 0);
1554 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1555 pmap_inval_pde(&pmap->pm_pdir[pdirindex], pmap,
1556 (vm_offset_t)pdirindex << SEG_SHIFT);
1561 * Weed out invalid mappings. Note: we assume that the page
1562 * directory table is always allocated, and in kernel virtual.
1568 * Limit our scan to either the end of the va represented
1569 * by the current page table page, or to the end of the
1570 * range being removed.
1576 * NOTE: pmap_remove_pte() can block.
1578 for (; sindex != pdnxt; sindex++) {
1581 ptbase = get_ptbase(pmap, sindex << PAGE_SHIFT);
1584 va = i386_ptob(sindex);
1585 if (pmap_remove_pte(pmap, ptbase, va))
1589 lwkt_reltoken(&vm_token);
1590 vm_object_drop(pmap->pm_pteobj);
1594 * Removes this physical page from all physical maps in which it resides.
1595 * Reflects back modify bits to the pager.
1597 * This routine may not be called from an interrupt.
1602 pmap_remove_all(vm_page_t m)
1607 #if defined(PMAP_DIAGNOSTIC)
1609 * XXX this makes pmap_page_protect(NONE) illegal for non-managed
1612 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
1613 panic("pmap_page_protect: illegal for unmanaged page, va: 0x%08llx", (long long)VM_PAGE_TO_PHYS(m));
1617 lwkt_gettoken(&vm_token);
1618 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
1619 KKASSERT(pv->pv_pmap->pm_stats.resident_count > 0);
1620 --pv->pv_pmap->pm_stats.resident_count;
1622 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
1623 KKASSERT(pte != NULL);
1625 tpte = pmap_inval_loadandclear(pte, pv->pv_pmap, pv->pv_va);
1626 if (tpte & VPTE_WIRED)
1627 --pv->pv_pmap->pm_stats.wired_count;
1628 KKASSERT(pv->pv_pmap->pm_stats.wired_count >= 0);
1631 vm_page_flag_set(m, PG_REFERENCED);
1634 * Update the vm_page_t clean and reference bits.
1636 if (tpte & VPTE_M) {
1637 #if defined(PMAP_DIAGNOSTIC)
1638 if (pmap_nw_modified((pt_entry_t) tpte)) {
1640 "pmap_remove_all: modified page not writable: va: 0x%x, pte: 0x%x\n",
1644 if (pmap_track_modified(pv->pv_pmap, pv->pv_va))
1647 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1648 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
1649 ++pv->pv_pmap->pm_generation;
1650 m->md.pv_list_count--;
1651 atomic_add_int(&m->object->agg_pv_list_count, -1);
1652 if (TAILQ_EMPTY(&m->md.pv_list))
1653 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1654 vm_object_hold(pv->pv_pmap->pm_pteobj);
1655 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem);
1656 vm_object_drop(pv->pv_pmap->pm_pteobj);
1659 KKASSERT((m->flags & (PG_MAPPED | PG_WRITEABLE)) == 0);
1660 lwkt_reltoken(&vm_token);
1664 * Set the physical protection on the specified range of this map
1667 * This function may not be called from an interrupt if the map is
1668 * not the kernel_pmap.
1673 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
1677 vm_offset_t pdnxt, ptpaddr;
1678 vm_pindex_t sindex, eindex;
1684 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
1685 pmap_remove(pmap, sva, eva);
1689 if (prot & VM_PROT_WRITE)
1692 lwkt_gettoken(&vm_token);
1693 ptbase = get_ptbase(pmap, sva);
1695 sindex = (sva >> PAGE_SHIFT);
1696 eindex = (eva >> PAGE_SHIFT);
1699 for (; sindex < eindex; sindex = pdnxt) {
1703 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1705 pdirindex = sindex / NPDEPG;
1708 * Clear the modified and writable bits for a 4m page.
1709 * Throw away the modified bit (?)
1711 if (((ptpaddr = pmap->pm_pdir[pdirindex]) & VPTE_PS) != 0) {
1712 pmap_clean_pde(&pmap->pm_pdir[pdirindex], pmap,
1713 (vm_offset_t)pdirindex << SEG_SHIFT);
1714 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1719 * Weed out invalid mappings. Note: we assume that the page
1720 * directory table is always allocated, and in kernel virtual.
1725 if (pdnxt > eindex) {
1729 for (; sindex != pdnxt; sindex++) {
1734 * Clean managed pages and also check the accessed
1735 * bit. Just remove write perms for unmanaged
1736 * pages. Be careful of races, turning off write
1737 * access will force a fault rather then setting
1738 * the modified bit at an unexpected time.
1740 ptep = &ptbase[sindex - sbase];
1741 if (*ptep & VPTE_MANAGED) {
1742 pbits = pmap_clean_pte(ptep, pmap,
1745 if (pbits & VPTE_A) {
1746 m = PHYS_TO_VM_PAGE(pbits);
1747 vm_page_flag_set(m, PG_REFERENCED);
1748 atomic_clear_long(ptep, VPTE_A);
1750 if (pbits & VPTE_M) {
1751 if (pmap_track_modified(pmap, i386_ptob(sindex))) {
1753 m = PHYS_TO_VM_PAGE(pbits);
1758 pbits = pmap_setro_pte(ptep, pmap,
1763 lwkt_reltoken(&vm_token);
1767 * Enter a managed page into a pmap. If the page is not wired related pmap
1768 * data can be destroyed at any time for later demand-operation.
1770 * Insert the vm_page (m) at virtual address (v) in (pmap), with the
1771 * specified protection, and wire the mapping if requested.
1773 * NOTE: This routine may not lazy-evaluate or lose information. The
1774 * page must actually be inserted into the given map NOW.
1776 * NOTE: When entering a page at a KVA address, the pmap must be the
1782 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1788 vpte_t origpte, newpte;
1796 vm_object_hold(pmap->pm_pteobj);
1797 lwkt_gettoken(&vm_token);
1800 * Get the page table page. The kernel_pmap's page table pages
1801 * are preallocated and have no associated vm_page_t.
1803 if (pmap == &kernel_pmap)
1806 mpte = pmap_allocpte(pmap, va);
1808 pte = pmap_pte(pmap, va);
1811 * Page Directory table entry not valid, we need a new PT page
1812 * and pmap_allocpte() didn't give us one. Oops!
1815 panic("pmap_enter: invalid page directory pmap=%p, va=0x%p\n",
1820 * Deal with races on the original mapping (though don't worry
1821 * about VPTE_A races) by cleaning it. This will force a fault
1822 * if an attempt is made to write to the page.
1824 pa = VM_PAGE_TO_PHYS(m) & VPTE_FRAME;
1825 origpte = pmap_clean_pte(pte, pmap, va);
1826 opa = origpte & VPTE_FRAME;
1828 if (origpte & VPTE_PS)
1829 panic("pmap_enter: attempted pmap_enter on 4MB page");
1832 * Mapping has not changed, must be protection or wiring change.
1834 if (origpte && (opa == pa)) {
1836 * Wiring change, just update stats. We don't worry about
1837 * wiring PT pages as they remain resident as long as there
1838 * are valid mappings in them. Hence, if a user page is wired,
1839 * the PT page will be also.
1841 if (wired && ((origpte & VPTE_WIRED) == 0))
1842 ++pmap->pm_stats.wired_count;
1843 else if (!wired && (origpte & VPTE_WIRED))
1844 --pmap->pm_stats.wired_count;
1845 KKASSERT(pmap->pm_stats.wired_count >= 0);
1848 * Remove the extra pte reference. Note that we cannot
1849 * optimize the RO->RW case because we have adjusted the
1850 * wiring count above and may need to adjust the wiring
1857 * We might be turning off write access to the page,
1858 * so we go ahead and sense modify status.
1860 if (origpte & VPTE_MANAGED) {
1861 if ((origpte & VPTE_M) &&
1862 pmap_track_modified(pmap, va)) {
1864 om = PHYS_TO_VM_PAGE(opa);
1868 KKASSERT(m->flags & PG_MAPPED);
1873 * Mapping has changed, invalidate old range and fall through to
1874 * handle validating new mapping.
1878 err = pmap_remove_pte(pmap, pte, va);
1880 panic("pmap_enter: pte vanished, va: %p", (void *)va);
1881 pte = pmap_pte(pmap, va);
1882 origpte = pmap_clean_pte(pte, pmap, va);
1883 opa = origpte & VPTE_FRAME;
1885 kprintf("pmap_enter: Warning, raced pmap %p va %p\n",
1891 * Enter on the PV list if part of our managed memory. Note that we
1892 * raise IPL while manipulating pv_table since pmap_enter can be
1893 * called at interrupt time.
1895 if (pmap_initialized &&
1896 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
1897 pmap_insert_entry(pmap, va, mpte, m);
1899 vm_page_flag_set(m, PG_MAPPED);
1903 * Increment counters
1905 ++pmap->pm_stats.resident_count;
1907 pmap->pm_stats.wired_count++;
1911 * Now validate mapping with desired protection/wiring.
1913 newpte = (vm_offset_t) (pa | pte_prot(pmap, prot) | VPTE_V);
1916 newpte |= VPTE_WIRED;
1917 if (pmap != &kernel_pmap)
1921 * If the mapping or permission bits are different from the
1922 * (now cleaned) original pte, an update is needed. We've
1923 * already downgraded or invalidated the page so all we have
1924 * to do now is update the bits.
1926 * XXX should we synchronize RO->RW changes to avoid another
1929 if ((origpte & ~(VPTE_W|VPTE_M|VPTE_A)) != newpte) {
1930 *pte = newpte | VPTE_A;
1931 if (newpte & VPTE_W)
1932 vm_page_flag_set(m, PG_WRITEABLE);
1934 KKASSERT((newpte & VPTE_MANAGED) == 0 || m->flags & PG_MAPPED);
1935 lwkt_reltoken(&vm_token);
1936 vm_object_drop(pmap->pm_pteobj);
1940 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
1942 * Currently this routine may only be used on user pmaps, not kernel_pmap.
1945 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
1953 KKASSERT(pmap != &kernel_pmap);
1955 KKASSERT(va >= VM_MIN_USER_ADDRESS && va < VM_MAX_USER_ADDRESS);
1958 * Calculate pagetable page (mpte), allocating it if necessary.
1960 * A held page table page (mpte), or NULL, is passed onto the
1961 * section following.
1963 ptepindex = va >> PDRSHIFT;
1965 vm_object_hold(pmap->pm_pteobj);
1966 lwkt_gettoken(&vm_token);
1970 * Get the page directory entry
1972 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
1975 * If the page table page is mapped, we just increment
1976 * the hold count, and activate it.
1979 if (ptepa & VPTE_PS)
1980 panic("pmap_enter_quick: unexpected mapping into 4MB page");
1981 if (pmap->pm_ptphint &&
1982 (pmap->pm_ptphint->pindex == ptepindex)) {
1983 mpte = pmap->pm_ptphint;
1985 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1986 pmap->pm_ptphint = mpte;
1987 vm_page_wakeup(mpte);
1992 mpte = _pmap_allocpte(pmap, ptepindex);
1994 } while (mpte == NULL);
1997 * Ok, now that the page table page has been validated, get the pte.
1998 * If the pte is already mapped undo mpte's hold_count and
2001 pte = pmap_pte(pmap, va);
2003 pmap_unwire_pte_hold(pmap, mpte);
2004 lwkt_reltoken(&vm_token);
2005 vm_object_drop(pmap->pm_pteobj);
2010 * Enter on the PV list if part of our managed memory. Note that we
2011 * raise IPL while manipulating pv_table since pmap_enter can be
2012 * called at interrupt time.
2014 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2015 pmap_insert_entry(pmap, va, mpte, m);
2016 vm_page_flag_set(m, PG_MAPPED);
2020 * Increment counters
2022 ++pmap->pm_stats.resident_count;
2024 pa = VM_PAGE_TO_PHYS(m);
2027 * Now validate mapping with RO protection
2029 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
2030 *pte = (vpte_t)pa | VPTE_V | VPTE_U;
2032 *pte = (vpte_t)pa | VPTE_V | VPTE_U | VPTE_MANAGED;
2033 /*pmap_inval_add(&info, pmap, va); shouldn't be needed 0->valid */
2034 /*pmap_inval_flush(&info); don't need for vkernel */
2035 lwkt_reltoken(&vm_token);
2036 vm_object_drop(pmap->pm_pteobj);
2040 * Extract the physical address for the translation at the specified
2041 * virtual address in the pmap.
2043 * The caller must hold vm_token if non-blocking operation is desired.
2047 pmap_extract(pmap_t pmap, vm_offset_t va)
2052 lwkt_gettoken(&vm_token);
2053 if (pmap && (pte = pmap->pm_pdir[va >> SEG_SHIFT]) != 0) {
2054 if (pte & VPTE_PS) {
2055 rtval = pte & ~((vpte_t)(1 << SEG_SHIFT) - 1);
2056 rtval |= va & SEG_MASK;
2058 pte = *get_ptbase(pmap, va);
2059 rtval = (pte & VPTE_FRAME) | (va & PAGE_MASK);
2064 lwkt_reltoken(&vm_token);
2068 #define MAX_INIT_PT (96)
2071 * This routine preloads the ptes for a given object into the specified pmap.
2072 * This eliminates the blast of soft faults on process startup and
2073 * immediately after an mmap.
2077 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
2080 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
2081 vm_object_t object, vm_pindex_t pindex,
2082 vm_size_t size, int limit)
2084 struct rb_vm_page_scan_info info;
2089 * We can't preinit if read access isn't set or there is no pmap
2092 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
2096 * We can't preinit if the pmap is not the current pmap
2098 lp = curthread->td_lwp;
2099 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
2102 psize = size >> PAGE_SHIFT;
2104 if ((object->type != OBJT_VNODE) ||
2105 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
2106 (object->resident_page_count > MAX_INIT_PT))) {
2110 if (psize + pindex > object->size) {
2111 if (object->size < pindex)
2113 psize = object->size - pindex;
2120 * Use a red-black scan to traverse the requested range and load
2121 * any valid pages found into the pmap.
2123 * We cannot safely scan the object's memq unless we are in a
2124 * critical section since interrupts can remove pages from objects.
2126 info.start_pindex = pindex;
2127 info.end_pindex = pindex + psize - 1;
2133 vm_object_hold(object);
2134 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2135 pmap_object_init_pt_callback, &info);
2136 vm_object_drop(object);
2140 * The caller must hold vm_token.
2144 pmap_object_init_pt_callback(vm_page_t p, void *data)
2146 struct rb_vm_page_scan_info *info = data;
2147 vm_pindex_t rel_index;
2150 * don't allow an madvise to blow away our really
2151 * free pages allocating pv entries.
2153 if ((info->limit & MAP_PREFAULT_MADVISE) &&
2154 vmstats.v_free_count < vmstats.v_free_reserved) {
2157 if (vm_page_busy_try(p, TRUE))
2159 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2160 (p->flags & PG_FICTITIOUS) == 0) {
2161 if ((p->queue - p->pc) == PQ_CACHE)
2162 vm_page_deactivate(p);
2163 rel_index = p->pindex - info->start_pindex;
2164 pmap_enter_quick(info->pmap,
2165 info->addr + i386_ptob(rel_index), p);
2172 * Return TRUE if the pmap is in shape to trivially
2173 * pre-fault the specified address.
2175 * Returns FALSE if it would be non-trivial or if a
2176 * pte is already loaded into the slot.
2181 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
2186 lwkt_gettoken(&vm_token);
2187 if ((*pmap_pde(pmap, addr)) == 0) {
2190 pte = get_ptbase(pmap, addr);
2191 ret = (*pte) ? 0 : 1;
2193 lwkt_reltoken(&vm_token);
2198 * Change the wiring attribute for a map/virtual-address pair.
2199 * The mapping must already exist in the pmap.
2201 * No other requirements.
2204 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
2211 lwkt_gettoken(&vm_token);
2212 pte = get_ptbase(pmap, va);
2214 if (wired && (*pte & VPTE_WIRED) == 0)
2215 ++pmap->pm_stats.wired_count;
2216 else if (!wired && (*pte & VPTE_WIRED))
2217 --pmap->pm_stats.wired_count;
2218 KKASSERT(pmap->pm_stats.wired_count >= 0);
2221 * Wiring is not a hardware characteristic so there is no need to
2222 * invalidate TLB. However, in an SMP environment we must use
2223 * a locked bus cycle to update the pte (if we are not using
2224 * the pmap_inval_*() API that is)... it's ok to do this for simple
2228 atomic_set_long(pte, VPTE_WIRED);
2230 atomic_clear_long(pte, VPTE_WIRED);
2231 lwkt_reltoken(&vm_token);
2235 * Copy the range specified by src_addr/len
2236 * from the source map to the range dst_addr/len
2237 * in the destination map.
2239 * This routine is only advisory and need not do anything.
2242 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
2243 vm_size_t len, vm_offset_t src_addr)
2246 vm_offset_t end_addr = src_addr + len;
2253 * XXX BUGGY. Amoung other things srcmpte is assumed to remain
2254 * valid through blocking calls, and that's just not going to
2261 if (dst_addr != src_addr)
2263 if (dst_pmap->pm_pdir == NULL)
2265 if (src_pmap->pm_pdir == NULL)
2268 lwkt_gettoken(&vm_token);
2270 src_frame = get_ptbase1(src_pmap, src_addr);
2271 dst_frame = get_ptbase2(dst_pmap, src_addr);
2274 * critical section protection is required to maintain the page/object
2275 * association, interrupts can free pages and remove them from
2278 for (addr = src_addr; addr < end_addr; addr = pdnxt) {
2279 vpte_t *src_pte, *dst_pte;
2280 vm_page_t dstmpte, srcmpte;
2281 vm_offset_t srcptepaddr;
2284 if (addr >= VM_MAX_USER_ADDRESS)
2285 panic("pmap_copy: invalid to pmap_copy page tables\n");
2288 * Don't let optional prefaulting of pages make us go
2289 * way below the low water mark of free pages or way
2290 * above high water mark of used pv entries.
2292 if (vmstats.v_free_count < vmstats.v_free_reserved ||
2293 pv_entry_count > pv_entry_high_water)
2296 pdnxt = ((addr + PAGE_SIZE*NPTEPG) & ~(PAGE_SIZE*NPTEPG - 1));
2297 ptepindex = addr >> PDRSHIFT;
2299 srcptepaddr = (vm_offset_t) src_pmap->pm_pdir[ptepindex];
2300 if (srcptepaddr == 0)
2303 if (srcptepaddr & VPTE_PS) {
2304 if (dst_pmap->pm_pdir[ptepindex] == 0) {
2305 dst_pmap->pm_pdir[ptepindex] = (vpte_t)srcptepaddr;
2306 dst_pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE;
2311 srcmpte = vm_page_lookup(src_pmap->pm_pteobj, ptepindex);
2312 if ((srcmpte == NULL) || (srcmpte->hold_count == 0) ||
2313 (srcmpte->flags & PG_BUSY)) {
2317 if (pdnxt > end_addr)
2320 src_pte = src_frame + ((addr - src_addr) >> PAGE_SHIFT);
2321 dst_pte = dst_frame + ((addr - src_addr) >> PAGE_SHIFT);
2322 while (addr < pdnxt) {
2327 * we only virtual copy managed pages
2329 if ((ptetemp & VPTE_MANAGED) != 0) {
2331 * We have to check after allocpte for the
2332 * pte still being around... allocpte can
2335 * pmap_allocpte can block, unfortunately
2336 * we have to reload the tables.
2338 dstmpte = pmap_allocpte(dst_pmap, addr);
2339 src_frame = get_ptbase1(src_pmap, src_addr);
2340 dst_frame = get_ptbase2(dst_pmap, src_addr);
2342 if ((*dst_pte == 0) && (ptetemp = *src_pte) &&
2343 (ptetemp & VPTE_MANAGED) != 0) {
2345 * Clear the modified and accessed
2346 * (referenced) bits during the copy.
2348 * We do not have to clear the write
2349 * bit to force a fault-on-modify
2350 * because the real kernel's target
2351 * pmap is empty and will fault anyway.
2353 m = PHYS_TO_VM_PAGE(ptetemp);
2354 *dst_pte = ptetemp & ~(VPTE_M | VPTE_A);
2355 ++dst_pmap->pm_stats.resident_count;
2356 pmap_insert_entry(dst_pmap, addr,
2358 KKASSERT(m->flags & PG_MAPPED);
2360 pmap_unwire_pte_hold(dst_pmap, dstmpte);
2362 if (dstmpte->hold_count >= srcmpte->hold_count)
2370 lwkt_reltoken(&vm_token);
2376 * Zero the specified PA by mapping the page into KVM and clearing its
2379 * This function may be called from an interrupt and no locking is
2383 pmap_zero_page(vm_paddr_t phys)
2385 struct mdglobaldata *gd = mdcpu;
2389 panic("pmap_zero_page: CMAP3 busy");
2390 *gd->gd_CMAP3 = VPTE_V | VPTE_R | VPTE_W | (phys & VPTE_FRAME) | VPTE_A | VPTE_M;
2391 madvise(gd->gd_CADDR3, PAGE_SIZE, MADV_INVAL);
2393 bzero(gd->gd_CADDR3, PAGE_SIZE);
2399 * pmap_page_assertzero:
2401 * Assert that a page is empty, panic if it isn't.
2404 pmap_page_assertzero(vm_paddr_t phys)
2406 struct mdglobaldata *gd = mdcpu;
2411 panic("pmap_zero_page: CMAP3 busy");
2412 *gd->gd_CMAP3 = VPTE_V | VPTE_R | VPTE_W |
2413 (phys & VPTE_FRAME) | VPTE_A | VPTE_M;
2414 madvise(gd->gd_CADDR3, PAGE_SIZE, MADV_INVAL);
2415 for (i = 0; i < PAGE_SIZE; i += 4) {
2416 if (*(int *)((char *)gd->gd_CADDR3 + i) != 0) {
2417 panic("pmap_page_assertzero() @ %p not zero!\n",
2418 (void *)gd->gd_CADDR3);
2428 * Zero part of a physical page by mapping it into memory and clearing
2429 * its contents with bzero.
2431 * off and size may not cover an area beyond a single hardware page.
2434 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
2436 struct mdglobaldata *gd = mdcpu;
2440 panic("pmap_zero_page: CMAP3 busy");
2441 *gd->gd_CMAP3 = VPTE_V | VPTE_R | VPTE_W |
2442 (phys & VPTE_FRAME) | VPTE_A | VPTE_M;
2443 madvise(gd->gd_CADDR3, PAGE_SIZE, MADV_INVAL);
2445 bzero((char *)gd->gd_CADDR3 + off, size);
2453 * Copy the physical page from the source PA to the target PA.
2454 * This function may be called from an interrupt. No locking
2458 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
2460 struct mdglobaldata *gd = mdcpu;
2463 if (*(int *) gd->gd_CMAP1)
2464 panic("pmap_copy_page: CMAP1 busy");
2465 if (*(int *) gd->gd_CMAP2)
2466 panic("pmap_copy_page: CMAP2 busy");
2468 *(int *) gd->gd_CMAP1 = VPTE_V | VPTE_R | (src & PG_FRAME) | VPTE_A;
2469 *(int *) gd->gd_CMAP2 = VPTE_V | VPTE_R | VPTE_W | (dst & VPTE_FRAME) | VPTE_A | VPTE_M;
2471 madvise(gd->gd_CADDR1, PAGE_SIZE, MADV_INVAL);
2472 madvise(gd->gd_CADDR2, PAGE_SIZE, MADV_INVAL);
2474 bcopy(gd->gd_CADDR1, gd->gd_CADDR2, PAGE_SIZE);
2476 *(int *) gd->gd_CMAP1 = 0;
2477 *(int *) gd->gd_CMAP2 = 0;
2482 * pmap_copy_page_frag:
2484 * Copy the physical page from the source PA to the target PA.
2485 * This function may be called from an interrupt. No locking
2489 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
2491 struct mdglobaldata *gd = mdcpu;
2494 if (*(int *) gd->gd_CMAP1)
2495 panic("pmap_copy_page: CMAP1 busy");
2496 if (*(int *) gd->gd_CMAP2)
2497 panic("pmap_copy_page: CMAP2 busy");
2499 *(int *) gd->gd_CMAP1 = VPTE_V | (src & VPTE_FRAME) | VPTE_A;
2500 *(int *) gd->gd_CMAP2 = VPTE_V | VPTE_R | VPTE_W | (dst & VPTE_FRAME) | VPTE_A | VPTE_M;
2502 madvise(gd->gd_CADDR1, PAGE_SIZE, MADV_INVAL);
2503 madvise(gd->gd_CADDR2, PAGE_SIZE, MADV_INVAL);
2505 bcopy((char *)gd->gd_CADDR1 + (src & PAGE_MASK),
2506 (char *)gd->gd_CADDR2 + (dst & PAGE_MASK),
2509 *(int *) gd->gd_CMAP1 = 0;
2510 *(int *) gd->gd_CMAP2 = 0;
2515 * Returns true if the pmap's pv is one of the first
2516 * 16 pvs linked to from this page. This count may
2517 * be changed upwards or downwards in the future; it
2518 * is only necessary that true be returned for a small
2519 * subset of pmaps for proper page aging.
2524 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
2529 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2533 lwkt_gettoken(&vm_token);
2535 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2536 if (pv->pv_pmap == pmap) {
2537 lwkt_reltoken(&vm_token);
2545 lwkt_reltoken(&vm_token);
2551 * Remove all pages from specified address space
2552 * this aids process exit speeds. Also, this code
2553 * is special cased for current process only, but
2554 * can have the more generic (and slightly slower)
2555 * mode enabled. This is much faster than pmap_remove
2556 * in the case of running down an entire address space.
2561 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
2566 int32_t save_generation;
2568 if (pmap->pm_pteobj)
2569 vm_object_hold(pmap->pm_pteobj);
2570 lwkt_gettoken(&vm_token);
2571 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
2572 if (pv->pv_va >= eva || pv->pv_va < sva) {
2573 npv = TAILQ_NEXT(pv, pv_plist);
2577 KKASSERT(pmap == pv->pv_pmap);
2579 pte = pmap_pte(pmap, pv->pv_va);
2582 * We cannot remove wired pages from a process' mapping
2585 if (*pte & VPTE_WIRED) {
2586 npv = TAILQ_NEXT(pv, pv_plist);
2589 tpte = pmap_inval_loadandclear(pte, pmap, pv->pv_va);
2591 m = PHYS_TO_VM_PAGE(tpte);
2593 KASSERT(m < &vm_page_array[vm_page_array_size],
2594 ("pmap_remove_pages: bad tpte %lx", tpte));
2596 KKASSERT(pmap->pm_stats.resident_count > 0);
2597 --pmap->pm_stats.resident_count;
2600 * Update the vm_page_t clean and reference bits.
2602 if (tpte & VPTE_M) {
2606 npv = TAILQ_NEXT(pv, pv_plist);
2607 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
2608 save_generation = ++pmap->pm_generation;
2610 m->md.pv_list_count--;
2611 atomic_add_int(&m->object->agg_pv_list_count, -1);
2612 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2613 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
2614 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2616 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem);
2620 * Restart the scan if we blocked during the unuse or free
2621 * calls and other removals were made.
2623 if (save_generation != pmap->pm_generation) {
2624 kprintf("Warning: pmap_remove_pages race-A avoided\n");
2625 npv = TAILQ_FIRST(&pmap->pm_pvlist);
2628 lwkt_reltoken(&vm_token);
2629 if (pmap->pm_pteobj)
2630 vm_object_drop(pmap->pm_pteobj);
2634 * pmap_testbit tests bits in active mappings of a VM page.
2636 * The caller must hold vm_token
2639 pmap_testbit(vm_page_t m, int bit)
2644 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2647 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
2652 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2654 * if the bit being tested is the modified bit, then
2655 * mark clean_map and ptes as never
2658 if (bit & (VPTE_A|VPTE_M)) {
2659 if (!pmap_track_modified(pv->pv_pmap, pv->pv_va))
2663 #if defined(PMAP_DIAGNOSTIC)
2665 kprintf("Null pmap (tb) at va: 0x%x\n", pv->pv_va);
2669 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
2680 * This routine is used to clear bits in ptes. Certain bits require special
2681 * handling, in particular (on virtual kernels) the VPTE_M (modify) bit.
2683 * This routine is only called with certain VPTE_* bit combinations.
2685 * The caller must hold vm_token
2687 static __inline void
2688 pmap_clearbit(vm_page_t m, int bit)
2694 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2700 * Loop over all current mappings setting/clearing as appropos If
2701 * setting RO do we need to clear the VAC?
2703 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2705 * don't write protect pager mappings
2707 if (bit == VPTE_W) {
2708 if (!pmap_track_modified(pv->pv_pmap, pv->pv_va))
2712 #if defined(PMAP_DIAGNOSTIC)
2714 kprintf("Null pmap (cb) at va: 0x%x\n", pv->pv_va);
2720 * Careful here. We can use a locked bus instruction to
2721 * clear VPTE_A or VPTE_M safely but we need to synchronize
2722 * with the target cpus when we mess with VPTE_W.
2724 * On virtual kernels we must force a new fault-on-write
2725 * in the real kernel if we clear the Modify bit ourselves,
2726 * otherwise the real kernel will not get a new fault and
2727 * will never set our Modify bit again.
2729 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
2731 if (bit == VPTE_W) {
2733 * We must also clear VPTE_M when clearing
2736 pbits = pmap_clean_pte(pte, pv->pv_pmap,
2740 } else if (bit == VPTE_M) {
2742 * We do not have to make the page read-only
2743 * when clearing the Modify bit. The real
2744 * kernel will make the real PTE read-only
2745 * or otherwise detect the write and set
2746 * our VPTE_M again simply by us invalidating
2747 * the real kernel VA for the pmap (as we did
2748 * above). This allows the real kernel to
2749 * handle the write fault without forwarding
2752 atomic_clear_long(pte, VPTE_M);
2753 } else if ((bit & (VPTE_W|VPTE_M)) == (VPTE_W|VPTE_M)) {
2755 * We've been asked to clear W & M, I guess
2756 * the caller doesn't want us to update
2757 * the dirty status of the VM page.
2759 pmap_clean_pte(pte, pv->pv_pmap, pv->pv_va);
2762 * We've been asked to clear bits that do
2763 * not interact with hardware.
2765 atomic_clear_long(pte, bit);
2773 * Lower the permission for all mappings to a given page.
2778 pmap_page_protect(vm_page_t m, vm_prot_t prot)
2780 if ((prot & VM_PROT_WRITE) == 0) {
2781 lwkt_gettoken(&vm_token);
2782 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
2783 pmap_clearbit(m, VPTE_W);
2784 vm_page_flag_clear(m, PG_WRITEABLE);
2788 lwkt_reltoken(&vm_token);
2793 pmap_phys_address(vm_pindex_t ppn)
2795 return (i386_ptob(ppn));
2799 * Return a count of reference bits for a page, clearing those bits.
2800 * It is not necessary for every reference bit to be cleared, but it
2801 * is necessary that 0 only be returned when there are truly no
2802 * reference bits set.
2804 * XXX: The exact number of bits to check and clear is a matter that
2805 * should be tested and standardized at some point in the future for
2806 * optimal aging of shared pages.
2811 pmap_ts_referenced(vm_page_t m)
2813 pv_entry_t pv, pvf, pvn;
2817 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2821 lwkt_gettoken(&vm_token);
2823 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
2828 pvn = TAILQ_NEXT(pv, pv_list);
2830 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2832 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2834 if (!pmap_track_modified(pv->pv_pmap, pv->pv_va))
2837 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
2839 if (pte && (*pte & VPTE_A)) {
2841 atomic_clear_long(pte, VPTE_A);
2843 atomic_clear_long_nonlocked(pte, VPTE_A);
2850 } while ((pv = pvn) != NULL && pv != pvf);
2852 lwkt_reltoken(&vm_token);
2859 * Return whether or not the specified physical page was modified
2860 * in any physical maps.
2865 pmap_is_modified(vm_page_t m)
2869 lwkt_gettoken(&vm_token);
2870 res = pmap_testbit(m, VPTE_M);
2871 lwkt_reltoken(&vm_token);
2876 * Clear the modify bits on the specified physical page.
2881 pmap_clear_modify(vm_page_t m)
2883 lwkt_gettoken(&vm_token);
2884 pmap_clearbit(m, VPTE_M);
2885 lwkt_reltoken(&vm_token);
2889 * Clear the reference bit on the specified physical page.
2894 pmap_clear_reference(vm_page_t m)
2896 lwkt_gettoken(&vm_token);
2897 pmap_clearbit(m, VPTE_A);
2898 lwkt_reltoken(&vm_token);
2902 * Miscellaneous support routines follow
2906 i386_protection_init(void)
2910 kp = protection_codes;
2911 for (prot = 0; prot < 8; prot++) {
2912 if (prot & VM_PROT_READ)
2914 if (prot & VM_PROT_WRITE)
2916 if (prot & VM_PROT_EXECUTE)
2925 * Map a set of physical memory pages into the kernel virtual
2926 * address space. Return a pointer to where it is mapped. This
2927 * routine is intended to be used for mapping device memory,
2930 * NOTE: we can't use pgeflag unless we invalidate the pages one at
2934 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
2936 vm_offset_t va, tmpva, offset;
2939 offset = pa & PAGE_MASK;
2940 size = roundup(offset + size, PAGE_SIZE);
2942 va = kmem_alloc_nofault(&kernel_map, size, PAGE_SIZE);
2944 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
2946 pa = pa & VPTE_FRAME;
2947 for (tmpva = va; size > 0;) {
2948 pte = KernelPTA + (tmpva >> PAGE_SHIFT);
2949 *pte = pa | VPTE_R | VPTE_W | VPTE_V; /* | pgeflag; */
2957 return ((void *)(va + offset));
2961 pmap_unmapdev(vm_offset_t va, vm_size_t size)
2963 vm_offset_t base, offset;
2965 base = va & VPTE_FRAME;
2966 offset = va & PAGE_MASK;
2967 size = roundup(offset + size, PAGE_SIZE);
2968 pmap_qremove(va, size >> PAGE_SHIFT);
2969 kmem_free(&kernel_map, base, size);
2975 * Perform the pmap work for mincore
2980 pmap_mincore(pmap_t pmap, vm_offset_t addr)
2986 lwkt_gettoken(&vm_token);
2988 ptep = pmap_pte(pmap, addr);
2990 lwkt_reltoken(&vm_token);
2994 if ((pte = *ptep) != 0) {
2997 val = MINCORE_INCORE;
2998 if ((pte & VPTE_MANAGED) == 0)
3001 pa = pte & VPTE_FRAME;
3003 m = PHYS_TO_VM_PAGE(pa);
3009 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
3011 * Modified by someone
3013 else if (m->dirty || pmap_is_modified(m))
3014 val |= MINCORE_MODIFIED_OTHER;
3019 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
3022 * Referenced by someone
3024 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
3025 val |= MINCORE_REFERENCED_OTHER;
3026 vm_page_flag_set(m, PG_REFERENCED);
3030 lwkt_reltoken(&vm_token);
3035 * Caller must hold vmspace->vm_map.token for oldvm and newvm
3038 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
3040 struct vmspace *oldvm;
3043 oldvm = p->p_vmspace;
3045 if (oldvm != newvm) {
3046 p->p_vmspace = newvm;
3047 KKASSERT(p->p_nthreads == 1);
3048 lp = RB_ROOT(&p->p_lwp_tree);
3049 pmap_setlwpvm(lp, newvm);
3051 sysref_get(&newvm->vm_sysref);
3052 sysref_put(&oldvm->vm_sysref);
3059 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
3061 struct vmspace *oldvm;
3065 oldvm = lp->lwp_vmspace;
3067 if (oldvm != newvm) {
3068 lp->lwp_vmspace = newvm;
3069 if (curthread->td_lwp == lp) {
3070 pmap = vmspace_pmap(newvm);
3072 atomic_set_cpumask(&pmap->pm_active, mycpu->gd_cpumask);
3074 pmap->pm_active |= 1;
3076 #if defined(SWTCH_OPTIM_STATS)
3079 pmap = vmspace_pmap(oldvm);
3081 atomic_clear_cpumask(&pmap->pm_active, mycpu->gd_cpumask);
3083 pmap->pm_active &= ~(cpumask_t)1;
3092 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
3095 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
3099 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
3104 * Used by kmalloc/kfree, page already exists at va
3107 pmap_kvtom(vm_offset_t va)
3111 KKASSERT(va >= KvaStart && va < KvaEnd);
3112 ptep = KernelPTA + (va >> PAGE_SHIFT);
3113 return(PHYS_TO_VM_PAGE(*ptep & PG_FRAME));